Demucs is an open-source music source separation model, published in the Facebook Research (Meta) repository on GitHub. Source separation means splitting a mixed recording into its parts. Version 4, Hybrid Transformer Demucs (htdemucs), works on both the waveform and the spectrogram of the audio and uses a cross-domain Transformer, a neural network that relates the parts of a sequence through attention.
A user installs the Python package with pip and runs the demucs command on an audio file. The result is one stereo WAV file at 44.1 kHz for each stem: drums, bass, vocals and other; MP3 output is an option. Flags choose the model, for example a fine-tuned or a smaller quantized one, or ask for two stems only, such as vocals and the rest. An experimental six-source model adds guitar and piano. Separation runs on a GPU or on the CPU, and the repository also has the code to train new models.
Demucs is free under the MIT license. Its author has left Meta, so this repository is no longer maintained; his fork at github.com/adefossez/demucs receives only important bug fixes.